Kubernetes Job cleanup leaks headless Services

[ethanbates-ez]

Background

The Kubernetes Job launcher creates a headless Service to provide stable DNS names for indexed Pods. Job cleanup deletes only the Job; the separate Service has no ownerReferences and survives indefinitely with its generated EndpointSlice. Repeated executions therefore accumulate Kubernetes API objects until cluster-wide informer caches can exhaust memory. This reproduces on Kubernetes 1.34 for Job-backed executions.

Acceptance Criteria

• Headless Services are lifecycle-owned by their Kubernetes Jobs.
• Cleanup removes Services created before ownership metadata existed.
• Failed Service creation does not leave a runnable Job.

Additional Context

Reproduction

1. Launch a Job-backed container execution.
2. Allow the execution to complete and run launcher cleanup.
3. Observe that the Job is deleted but its same-named headless Service remains.

Expected behavior: deleting or cleaning up the Job also removes the headless Service and generated EndpointSlice.

Proposed fix: #368